Reproducible Machine Learning for Banana Postharvest Quality Assessment: Data Provenance, Statistical Inference, and Accuracy Efficiency Trade-Offs

Reported accuracies for banana ripeness recognition are high but rarely comparable, because public releases mix source images with augmented exports and results are usually given as point estimates. This study proposes a reproducible, task-separated benchmark for visual banana postharvest quality assessment. Eight classifiers, including the compact MicroNet_Banana and lightweight and transformer-based models, were evaluated on the official 562-image test split of a classification release, and six detectors (five YOLO variants and RT-DETR-L) on the official 757-image test split of a detection release. Compared with previous work on the same release, the framework adds a provenance audit, hash-based split-independence checks, paired McNemar and equivalence tests with Holm adjustment, a standardized efficiency protocol with Pareto analysis, and external and cross-source validation. EfficientNet-B0 achieved the highest accuracy (99.47%; 95% CI: 98.44–99.82%) and was practically equivalent to three other lightweight models, whereas MicroNet_Banana reached 94.66% with 0.41 M parameters and 7.65 ms latency. The YOLO detectors (94.85–95.36% mAP@50) were practically equivalent, whereas RT-DETR-L was lower (93.36%). Removing near-duplicate test images changed the accuracy by less than 0.9 percentage points, but accuracy on an independent dataset fell to 32.8–61.7%. Model selection should therefore jointly consider accuracy, uncertainty, data provenance, generalization, and computational cost.